Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 14, 2026CHENNAI

The Defender's Dilemma: Measuring AI Refusal on Real Incident Response Artifacts

Rahul Kumar · Team Rahul

Submitted to AI Incident Response Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: The Defender's Dilemma: Measuring AI Refusal on Real Incident Response Artifacts

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

We tested whether AI models can help investigate AI-caused security incidents, or whether their own safety guardrails get in the way. Using 7 forensic analysis tasks built from real, publicly verified artifacts of the July 2026 Hugging Face intrusion including exploit code, infrastructure logs, and behavioral evidence, we tested 5 current models across neutral and incident-response-authorized framings. Claude Fable 5.1 blocked 92.9% of requests (13/14) via an API-level content filter before the model could generate a response. The other four models blocked zero and completed every task, including GPT-6 Astra from the same provider whose agents caused the incident. We also found an authorization paradox: adding professional incident-response framing causes Claude to block a prompt that it answers under neutral framing. On the hardest task, cryptographic weakness analysis, only Astra identified all vulnerabilities; open-weight models understood the code but missed the security flaws. The model that produces the strongest forensic analysis is also the one that blocks defenders from using it.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. The paper mentions a simple and useful case study showing that Claude Fable 5.1's content_filter blocked the most number of incident response requests, while other models in the evaluation did not. This finding could be a valuable data point for defenders. The limitations in the paper are also well documented. I really liked the C5 test where the author tried different wordings to see what triggered the block. However, it is not clear if the author tried something similar on the tasks that were blocked both ways, like in A1. That one has no attack words in the code, so the question is whether the code itself is what triggers the block or the words around it. This could be one of the future topics to discuss and expand on

  2. The project provides a useful, focused benchmark of defensive analysis using public incident artifacts, with released responses and repeated checks of selected findings. Distinguishing API-level blocking from textual refusals is operationally valuable, and the framing-dependent blocking result deserves further investigation.

Cite this project

@misc{kumar2026defenders,
  title = {{The Defender's Dilemma: Measuring AI Refusal on Real Incident Response Artifacts}},
  author = {Rahul Kumar},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/the-defenders-dilemma-measuring-ai-refusal-on-real-incident-response-artifacts-ftid}},
  url = {https://apartresearch.com/sprints/projects/the-defenders-dilemma-measuring-ai-refusal-on-real-incident-response-artifacts-ftid}
}

More from AI Incident Response Sprint

All 233 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026